Beyond the Human Eye: The Emergence of AI in Radiology

Key Takeaways

---

In an era marked by rapid technological advancement, the ability of computers to ‘see’ is no longer a novelty but a pivotal development. This is the concept of computer vision, a subset of computer science that is gaining significant interest and investment. The global computer vision market was valued at $15 billion in 2022 and is projected to reach $82.1 billion by 2032. 

One field leveraging computer vision in healthcare is radiology. Of the FDA’s new list of 882 approved devices (published on October 19, 2023, and updated May 2024), 76% are dedicated to radiology. (Figure 1)  

The sheer amount of available medical data since 2020 has accelerated the integration of AI-based technology within the field. Since 2013, hospitals have generated an average of over 50 million gigabytes of data each year, now encompassing about 30% of the world’s data overall. The type of data available includes vast amounts of imaging data with which AI can be trained at an accelerated rate, making AI-based image analysis a closer reality. Examples of the use of computer vision within radiology include discerning normal imaging from anomalies, such as pneumothorax or malignant tumors.

The Exponential Growth of Medical Data

To put this massive growth into context, it is important to understand that more data is available now than at any other time in history. The American Clinical and Climatological Association presented a report illustrating that it took 3.5 years to double medical data in 2010 versus a mere 73 days in 2020. To contextualize this in human terms, a radiologist who was shown 500 images during a standard 12-hour shift in 2008 can now sift through over 50,000 images. That is a 100 times increase in the amount of images in a 15-year span, and as these models become increasingly more sophisticated, the amount of insights provided will only grow.

AI In Image Recognition: Beyond Human Capabilities

Computer vision started in an MIT lab in 1966, when researchers linked a camera to a computer in order to teach the computer how to decipher different colored candy bars; computer vision today is in an entirely different stratosphere. According to Matt Lungren, MD, at Stanford University, there are currently models that outperform some of the country’s top radiologists. In an experiment to demonstrate this point, Lungren showed 420 X-rays to four Stanford radiologists with the objective to independently annotate for likely pneumonia. Within a week of training the algorithm, it could identify pneumonia more accurately than physicians.

These results are not  an indicator of  human inadequacy but rather an illustration of the training capacity and efficiency of AI. In the aforementioned study, Lungren references the training that the deployed model underwent, processing 112,120 frontal-view chest X-ray images in numerous iterations – a feat that is simply unattainable for human radiologists.  Likewise, AI has shown improved efficiency in detecting fractures and in digital breast tomosynthesis (DBT).

Practical Applications of AI in Radiology

Integration of AI in radiology has significant implications on two main levels:

Individual Level:

AI can be directly integrated into diagnostic devices like Xrays, MRI, CT, scanners, and ultrasounds and assist individual physicians with their workflows. Deep learning networking models from the University of Singapore have been found to correctly identify pneumothorax (collapsed lung) at an average rate of 93%.

Some of the biggest issues plaguing physicians are lack of staffing and high work volume. Many hospitals and radiology centers have high patient throughput and, thus, a high volume of imaging studies to read. AI can help by triaging and prioritizing studies, thereby ensuring quick attention to critical cases and accelerating diagnosis for patients where time is of the essence. For less critical cases, AI could even make a first-pass attempt at reading the image followed by a thorough review by a physician who can adjust and update as necessary.

Departmental Level:

AI can automate routine tasks such as data entry and image annotation, freeing radiologists’ valuable time to spend on patient care tasks. For instance, AI systems can automatically pull up historical patient information so that it’s ready for review alongside new imaging, giving physicians the appropriate knowledge and context about a patient’s medical history while interpreting the study.

Additionally, AI-driven predictive analytics can keep track of patient imaging, demographics, chief complaints, and diagnoses, allowing for easier future data mining and review.

AI & Human Collaboration in Radiology

The idea of artificial intelligence within the field of radiology is not a blind, technical leap forward, but rather is starting to seem like a necessity. A 2017 study highlighted that errors and discrepancies in radiology are “uncomfortably common,” coming in at a 3-5% error rate day to day. This is not to say that radiologists are not good at their job. However, it is reasonable to consider that AI assistance could help. (Also important to note that the definition of error and discrepancy differs and must be taken into account when interpreting statistics)

Despite Geoffrey Hinton’s statement that “Radiologists should stop training,” this collaboration is not about replacing human expertise with machines. Instead, a much more realistic expectation is a symbiotic relationship between radiologists and machines working together to manage workloads, increase efficiency, and improve the accuracy of radiological reads.

Collaboration between machine and human could also help some of the practical issues in radiology. An ACR Intersociety Meeting convened in 2022 to discuss the pressing challenge of the aging healthcare workforce. The average age of a radiologist was reported to be 51 years old, and the discipline has a high turnover rate. AI tools can help bridge this gap by helping existing radiologists balance their workload and attracting younger radiologists due to a more streamlined and effective process in diagnostic settings. 

As artificial intelligence becomes more usable, efficient, and intelligible, so do the opportunities for us to learn from it. The use of AI in radiology can be likened to the adoption of autopilot in aviation. Just as autopilot hasn’t replaced pilots but instead augmented their capabilities and safety, a striking parallel seems evident in the world of radiology.

The Future of AI in Radiology: Opportunities and Challenges

Open-source AI models are a key driver of advancement within radiology, ushering in a new era of collaborative and transparent development. These models improve with use as they continuously learn from a diverse array of data, a vital feature for radiology. 

Work is already advancing internationally, illustrating the benefits of AI and radiology. A study in Sweden reviewed AI-assisted breast cancer screening and found it reduced radiologists’ workload by 44% without compromising accuracy. Similarly, Nature Magazine showcased that scientists have recently developed an AI tool that can diagnose multiple health conditions, from ocular diseases to Parkinson’s disease – all on the basis of a retina scan. 

Opportunities

The collaborative approach mentioned above wherein AI works with human physicians is especially useful in addressing global disparities in healthcare access. Approximately three to four billion individuals worldwide lack access to basic radiology services, and only 25% of people in low to middle-income countries (LMICs) have access to cancer diagnostic services. Moreover, a 2018 report stated 70% of global cancer deaths occurred in LMICs. AI can play the game-changing role of bridging these gaps.

AI in radiology can also expand access to care for patients in rural communities right here at home.

It’s also not hard to imagine that an AI radiology app will exist at some point in the future. In this scenario, anyone can use the app and phone camera to scan a body part and check for injury (e.g., scan an injured ankle, and the app can check for fractures). Much like how iPhones now can record an EKG reading that patients then show their doctors, AI-read phone-based radiology images could also be sent to primary care offices for review and confirmation.

Challenges

AI is not without its limitations, and computer vision is obviously not perfect. The models act based on context. Discerning nuanced differences in patient imaging takes a vast amount of data and machine learning to help ensure that studies are correctly read. For instance, image artifacts could be misread as nodules or benign calcifications as something more ominous. Incorrectly assuming the worst diagnosis could lead to further unnecessary testing; mistaking a real problem for something benign could delay adequate, potentially life-saving care. This makes it imperative that training data and inputs are accurate and that the AI models themselves interpret the data appropriately.

Despite the potential benefits, opinions in the industry about implementing AI in radiology are mixed. A 2021 survey investigated the sentiments of radiologists and radiology residents regarding AI’s future in their field, revealing a correlation between the level of expertise and the willingness to integrate AI into radiological work. Radiologists and residents with limited knowledge of AI tools and capabilities were more apprehensive, while experts and those with more experience viewed AI tools positively. Specifically, 48% of radiologists and residents have an open and proactive attitude toward AI, while 38% fear replacement by AI. These findings underscore the need to make AI knowledge more accessible to radiology professionals and to increase AI awareness and education within the field.

Final Thoughts

Whether one is apprehensive or optimistic about the implications of AI, there is mounting evidence that it can play a large role in moving the radiology field forward. With much of the world lacking resources or access altogether to the field of radiology and medical errors killing or seriously injuring scores of Americans each year – advancements in diagnostic accuracy and streamlining processes should be approached with optimism, albeit cautiously.

Regulations and monitoring of models already placed in hospitals are crucial.

Accurately evaluating AI systems is the critical first step toward generating radiology reports that are clinically useful and trustworthy

Pranav Rajpurkar(Assistant Professor of Biomedical Informatics, Harvard)

The future of AI in radiology holds great promise. With thoughtful implementation and oversight, AI stands to revolutionize the field and streamline workflows, augmenting the efficiency and accuracy of medical imaging.